Critical for OpenShift AI

Must:

• Add GPU per-hour rate to price list in OCP cost models. Probably useful for cloud instances too.
• Support GPUs in the various cases: dedicated (crawl), multi-instance (walk) and virtualized (run)
• Support nvidia GPUs

Should:

• At least on-premise, itemize cost of GPU, since it's a scarce resource (as compared to cost of CPU or memory or storage). Is it the moment to implement COST-3820? Research will be needed to determine whether this itemization is possible on AWS, Azure and GCP.

Could:

• Support AMD GPUs

Notes

• They would have to install the NVIDIA Operator to get the GPU metrics in Prometheus. See https://github.com/NVIDIA/dcgm-exporter. 
• https://docs.nvidia.com/datacenter/cloud-native/openshift/23.9.2/time-slicing-gpus-in-openshift.html 
• https://docs.openshift.com/container-platform/4.11/monitoring/nvidia-gpu-admin-dashboard.html 

Epic Design Document

Feature Brainstorming Document

Kruize Research Documents

• GPU Optimization
• Exporting Metrics